Speaker
Description
Edge computing in space is revolutionizing on-board data processing for Earth Observation (EO) satellites, enabling real-time analysis of optical and SAR imagery. GMV has explored novel FPGA-based AI acceleration solutions using the rad-tolerant Xilinx Versal AI Edge, leveraging Vitis AI workflows to map deep learning models for vessel detection and fire hotspot identification. Various Xilinx methodologies were tested, including DPU-based acceleration, HLS-generated AI functions, and custom FPGA modules. The final implementation employs Unify, integrating CNN layers onto both DPUs and FPGA logic, where non-DPU-compatible functions are efficiently mapped onto dedicated IP cores accessible via Unify software calls.
On-board edge computing is transforming Earth Observation (EO) by reducing data downlink requirements while enabling real-time insights. GMV developed an AI-accelerated processing pipeline leveraging the rad-tolerant Xilinx Versal AI Edge due to its lower power consumption and optimized DPU architecture (transitioning from the Versal AI Core), which better suits Deep Learning and Machine Learning workloads. Using Vitis AI workflows, CNNs are efficiently deployed onto DPUs and FPGA logic, ensuring high-performance inference in orbit.
A ground-onboard partitioning approach is implemented. On-board edge-computation includes a first-stage triage based on simple AI model processing thumbnails to discard irrelevant data (e.g., non-useful land-only images for vessel detection or cloud-covered scenes obstructing analysis). The core edge computing functions are based on a reduced, yet complex, AI model onboard that filters and processes data in real time, drastically minimizing the amount of imagery transmitted to Earth. Selected data patches are prioritized for further on-ground refinement with complete AI model. Additionally, the pipeline processes raw sensor data from L0 to L1b or L1c directly onboard, enhancing autonomy and mission efficiency.
This approach demonstrates the feasibility of deploying AI-driven EO applications in space, leveraging Versal AI Edge’s advanced AI acceleration while ensuring efficient power consumption and optimized FPGA resource allocation.
Affiliation of author(s)
GMV
| Track | Artificial Intelligence/Machine Learning |
|---|
